Temporality in relation with Discourse Structure

1
Temporality in relation with Discourse Structure

• Corina Forascu1, 2, Ionut Cristian Pistol1, Dan
  Cristea1, 3
• 1Faculty of Computer Science, University Al. I.
  Cuza of Iasi, Romania
• 2 Research Institute for Artificial Intelligence
  of the Romanian Academy, Bucharest
• 3 Institute for Computer Science, Romanian
  Academy, Iasi, Romania
• corinfor,ipistol,dcristea_at_info.uaic.ro

2
Outline

• Goal
• Theory
• Discourse Structure RST Veins Theory
• Temporal Information
• Experiment
• Data
• Methods
• Results evaluation
• Conclusions

3
Goal

• Further investigate the interconnection between
  temporality and discourse structure
• quantitative investigation of the dependencies
  between RST (actually, a variation of it VT) and
  the temporal relations
• if such a link exists, use it to
• reduce the human effort in manual annotation of
  temporal relations
• automatically annotate temporal relations
• improve discourse parsing

4
Outline

• Goal
• Theory
• Discourse Structure RST Veins Theory
• Temporal Information
• Experiment
• Data
• Methods
• Results evaluation
• Conclusions

5
Discourse Structure Fundamental assumptions in
VT

• An inter-unit reference is possible only when the
  two units are in a structural relation one with
  the other
• The nucleus-satellite distinction together with
  discourse structure gives indications on the
  range of referents to which an anaphor can be
  resolved

6
Veins
7
Veins
8
Veins
9
Veins
10
Veins
11
Veins
12
Veins
13
Veins
14
Veins
15
Veins
16
Veins
17
Veins
18
Veins
19
Veins
20
Veins
21
Veins
22
Veins
23
Veins
24
Veins
25
Veins
26
Veins
27
Veins
28
Veins
29
Veins
30
Veins
31
Veins and referentiality

• VT claim on the strong interconnection between
  discourse structure and referentiality has been
  proven experimentally
• This link has been applied to
• recuperate easier and saver referential relations
  when vein structure is known
• parse discourse, when referential relations are
  known

32
Veins and temporality

• See whether veins are related to temporal
  relations as well
• the way veins are defined raises the intuition
  that they must be
• if so, then they can be used to
• recuperate easier and saver temporal relations
  when discourse structure is known
• help discourse parsing, when temporal relations
  are known

33
Temporal Information

• TimeML 1.1. specification language for the
  event-structure of narrative texts
• Mark-up of
• Events EVENT, MAKEINSTANCE
• Temporal anchoring TIMEX3, SIGNAL
• Dependencies between events and times TLINK,
  ALINK, SLINK

34
Temporal Information
10/30/89
Kellogg is so anxious to turn around Corn Flakes
sales that it soon will begin selling boxes for
as little as 99 cents.
35
Temporal Information
10/30/89
ltEVENT eid"e206" class"I_STATE"gt
Kellogg is so anxiouse206 to turn around Corn
Flakes sales that it soon will begin selling
boxes for as little as 99 cents.
36
Temporal Information
10/30/89
ltEVENT eid"e32" class"OCCURRENCE"gt
Kellogg is so anxiouse206 to turne32 around Corn
Flakes sales that it soon will begin selling
boxes for as little as 99 cents.
37
Temporal Information
10/30/89
ltEVENT eid"e33" class"ASPECTUAL"gt
Kellogg is so anxiouse206 to turne32 around Corn
Flakes sales that it soon will begine33 selling
boxes for as little as 99 cents.
38
Temporal Information
10/30/89
ltEVENT eid"e34" class"OCCURRENCE"gt
Kellogg is so anxiouse206 to turne32 around Corn
Flakes sales that it soon will begine33
sellinge34 boxes for as little as 99 cents.
39
Temporal Information
10/30/89t192
ltTIMEX3 tid"t192" type"DATE" temporalFunction"f
alse" functionInDocument"CREATION_TIME"
value"1989-10-30"gt10/30/89lt/TIMEX3gt
Kellogg is so anxiouse206 to turne32 around Corn
Flakes sales that it soon will begine33
sellinge34 boxes for as little as 99 cents.
40
Temporal Information
10/30/89t192
ltTIMEX3 tid"t207" type"DATE" temporalFunction"t
rue" functionInDocument"NONE" value"FUTURE_REF"
anchorTimeID"t192"gt
Kellogg is so anxiouse206 to turne32 around Corn
Flakes sales that it soont207 will begine33
sellinge34 boxes for as little as 99 cents.
41
Temporal Information
10/30/89t192
ltSIGNAL sid"s31"gt
Kellogg is so anxiouse206 tos31 turne32 around
Corn Flakes sales that it soont207 will begine33
sellinge34 boxes for as little as 99 cents.
42
Temporal Information
10/30/89t192
Kellogg is so anxiouse206 tos31 turne32 around
Corn Flakes sales that it soont207 will begine33
sellinge34 boxes for as little as 99 cents.
ltMAKEINSTANCE aspect"NONE" eiid"ei2019"
tense"PRESENT" eventID"e206" /gt ltMAKEINSTANCE
aspect"NONE" eiid"ei2020" tense"NONE"
eventID"e32" /gt ltMAKEINSTANCE aspect"NONE"
eiid"ei2021" tense"FUTURE" eventID"e33" /gt
ltMAKEINSTANCE aspect"PROGRESSIVE" eiid"ei2022"
tense"NONE" eventID"e34" /gt
43
Temporal Information
10/30/89t192
Kellogg is so anxiouse206 tos31 turne32 around
Corn Flakes sales that it soont207 will begine33
sellinge34 boxes for as little as 99 cents.
ltTLINK relatedToTime"t192" eventInstanceID"ei201
9" relType"INCLUDES" /gt ltTLINK
relatedToEventInstance"ei2021"
eventInstanceID"ei2019" relType"BEFORE" /gt
ltTLINK relatedToTime"t207" eventInstanceID"ei20
21" relType"IS_INCLUDED" /gt ltTLINK
relatedToTime"t192" eventInstanceID"ei2021"
relType"AFTER" /gt
44
Temporal Information
10/30/89t192
Kellogg is so anxiouse206 tos31 turne32 around
Corn Flakes sales that it soont207 will begine33
sellinge34 boxes for as little as 99 cents.
ltTLINK relatedToTime"t192" eventInstanceID"ei201
9" relType"INCLUDES" /gt ltTLINK
relatedToEventInstance"ei2021"
eventInstanceID"ei2019" relType"BEFORE" /gt
ltTLINK relatedToTime"t207" eventInstanceID"ei20
21" relType"IS_INCLUDED" /gt ltTLINK
relatedToTime"t192" eventInstanceID"ei2021"
relType"AFTER" /gt
45
Temporal Information
10/30/89t192
Kellogg is so anxiouse206 tos31 turne32 around
Corn Flakes sales that it soont207 will begine33
sellinge34 boxes for as little as 99 cents.
ltTLINK relatedToTime"t192" eventInstanceID"ei201
9" relType"INCLUDES" /gt ltTLINK
relatedToEventInstance"ei2021"
eventInstanceID"ei2019" relType"BEFORE" /gt
ltTLINK relatedToTime"t207" eventInstanceID"ei20
21" relType"IS_INCLUDED" /gt ltTLINK
relatedToTime"t192" eventInstanceID"ei2021"
relType"AFTER" /gt
46
Temporal Information
10/30/89t192
Kellogg is so anxiouse206 tos31 turne32 around
Corn Flakes sales that it soont207 will begine33
sellinge34 boxes for as little as 99 cents.
ltTLINK relatedToTime"t192" eventInstanceID"ei201
9" relType"INCLUDES" /gt ltTLINK
relatedToEventInstance"ei2021"
eventInstanceID"ei2019" relType"BEFORE" /gt
ltTLINK relatedToTime"t207" eventInstanceID"ei20
21" relType"IS_INCLUDED" /gt ltTLINK
relatedToTime"t192" eventInstanceID"ei2021"
relType"AFTER" /gt
47
Temporal Information
10/30/89t192
Kellogg is so anxiouse206 tos31 turne32 around
Corn Flakes sales that it soont207 will begine33
sellinge34 boxes for as little as 99 cents.
ltSLINK signalID"s31" subordinatedEventInstance"e
i2020" eventInstanceID"ei2019" relType"MODAL"
/gt   ltSLINK signalID"s31" subordinatedEventInst
ance"ei2020" eventInstanceID"ei2021"
relType"MODAL" /gt
48
Temporal Information
10/30/89t192
Kellogg is so anxiouse206 tos31 turne32 around
Corn Flakes sales that it soont207 will begine33
sellinge34 boxes for as little as 99 cents.
ltSLINK signalID"s31" subordinatedEventInstance"e
i2020" eventInstanceID"ei2019" relType"MODAL"
/gt   ltSLINK signalID"s31" subordinatedEventInst
ance"ei2020" eventInstanceID"ei2021"
relType"MODAL" /gt
49
Temporal Information
10/30/89t192
Kellogg is so anxiouse206 tos31 turne32 around
Corn Flakes sales that it soont207 will begine33
sellinge34 boxes for as little as 99 cents.
ltALINK relatedToEventInstance"ei2022"
eventInstanceID"ei2021" relType"INITIATES" /gt
50
Outline

• Goal
• Theory
• Discourse Structure RST Veins Theory
• Temporal Information
• Experiment
• Data
• Methods
• Results evaluation
• Conclusions

51
Experimental data (1)
52
Experimental data (2)
53
Experiment method
RST Treebank Lisp-notation
TimeBank XML
XML-notation
Computed Veins
Merged structure RST veins temporal
information
54
Experiment method

• lt?xml version"1.0" encoding"ISO-8859-1" ?gt
• ltTimeML xmlnsxsi"http//www.w3.org/2001/XMLSchem
  a-instance" xsinoNamespaceSchemaLocation"TimeML1
  .1.xsd"gt
•   ltDOCNOgtWSJ891030-0089lt/DOCNOgt ltDDgt 891030lt/DDgt
  
•   ltANgt891030-0089.lt/ANgt
•   ltHLgtKellogg Pours Water on Plans to Build Big
  Cereal Plant _at_ ---- _at_ By Richard Gibson _at_ Staff
  Reporter of The Wall Street Journallt/HLgt ltDDgt
•   ltTIMEX3 tid"t192" type"DATE"
  temporalFunction"false" functionInDocument"CREAT
  ION_TIME" value"1989-10-30"gt10/30/89lt/TIMEX3gtlt/DD
  gt
•   ltSOgtWALL STREET JOURNAL (J)lt/SOgtltCOgtKlt/COgt
  ltINgtFOOD PRODUCTS (FOD)lt/INgt
• - ltTEXTgtltRootgt
• - ltrel ID"1001" nuc"no" h"8" vein"8"gtltrel
  ID"0" nuc"no" h"8" vein"8"gtltrel ID"1"
  nuc"yes" h"8" vein"8"gt
• - ltrel ID"2" nuc"yes" h"8" vein"8"gtltrel
  ID"3" nuc"yes" h"8" vein"8"gt
• - ltrel ID"4" nuc"yes" h"8" vein"8"gtltrel
  ID"5" nuc"yes" h"8" vein"8"gt
• ltrel ID"6" nuc"yes" h"8" vein"8"gt
• ltsgtltseg ID"7" CONTINUE"" nuc"no" h"7"
  vein"7,8"gt
•   ltw pos"DT"gtThislt/wgtltw pos","gthasntlt/wgtltw
  pos"BEN"gtbeenlt/wgt
•  ltENAMEX TYPE"ORGANIZATION"gtltw
  pos"IN"gtKellogglt/wgtltw pos"NN"gtColt/wgtltw
  pos"."gt.lt/wgt lt/ENAMEXgt
•   ltw pos"CD" /gtltw pos"NN"gtslt/wgtltw
  pos"NN"gtyearlt/wgtltw pos"."gt.lt/wgt
•   lt/seggtlt/sgt
• ltsgtltseg ID"8" CONTINUE"" nuc"yes" h"8"
  vein"7,8"gt
•   ltw pos"DT"gtThelt/wgtltw pos"NN"gtoat-branlt/wgt

Veins
TEs
55
Experiment results a) relations-on-the-veins
56
Experiment results b) annotation effort
evaluation

• Same/adjacent edus 67.18
• Long distance edus 32.82
• 83.42 on veins
• 16.58 not on veins

57
Experiment results c) types of temporal
relations
58
Experiment results d) types of discourse
relations

• Temporal relations NOT on veins
• 30 ATTRIBUTION
• "I don't think / Kellogg can get back to 40
  this year," / he said.
• Mr. Maxwell said / he wouldn't be surprised
• VEIN .. Kellogg can get back to 40 this
  year , / Mr. Maxwell said ..
• 32 SPAN

59
Discussion

• Expectations fulfilled to a great extend
• temporal links are mainly located on the veins
• there is a subset of temporal relations that
  behave better in correlation with discourse
  structure
• there is a subset of discourse relations that
  behave worse in relation with temporal relations

60
Discussion exceptions

• Mainly in connection with indirect speech (speech
  act events do not correlate with temporal
  relations)
• the class of verbs involving speech acts anchors
  temporal relations addressing co-terms outside
  veins
• the failings to find co-terms in time relations
  on the veins happen mainly around the ATTRIBUTION
  rhetorical relation, the one involving speech
  acts
• BUT the same bad correlation around indirect
  speech rhetorical relations has been noticed also
  in correlation with referentiality (Cristea et
  al., 2000)

61
Discussion exploitation

• A method to quicken the manual annotation of
  temporal relations in texts which include already
  an annotation of the structure
• an interface could reveal only the discourse
  units on the veins
• by pruning neighbouring discourse units, the
  context of time correlates is enlarged
• Automatic time tracking
• pruning method less units should be searched
  back until a time correlate is discovered
• Automatic discourse parsing
• score better, as structurally placed on identical
  veins, the discourse units displaying time
  relationships between them

62
Discussion exploitation
The oat-bran craze has cost the world's largest
cereal maker market share. The company's
president quit suddenly. And now Kellogg is
indefinitely suspending work on what was to
be a 1 billion cereal plant. The company said
it was delaying construction because of
current market conditions. But the Memphis,
Tenn., facility wasn't to begin turning out
product until 1993, so the decision may
reveal a more pessimistic long-term outlook as
well. Kellogg, which hasn't been as
successful in capitalizing on the public's
health-oriented desire for oat bran as rival
General Mills Inc., has been losing share in
the 6 billion ready-to-eat cereal market.
Kellogg's current share is believed to be
slightly under 40 while General Mills'
share is about 27. Led by its oat-based
Cheerios line, General Mills has gained an
estimated 2 share so far this year, mostly at
the expense of Kellogg. Each share point is
worth about 60 million in sales. Analysts say
much of Kellogg's erosion has been in such
core brands as Corn Flakes, Rice Krispies and
Frosted Flakes, which represent nearly
one-third of its sales volume. Kellogg is so
anxious to turn around Corn Flakes sales
that it soon will begin selling boxes for as
little as 99 cents, trade sources say.
"Cheerios and Honey Nut Cheerios have eaten away
sales normally going to Kellogg's corn-based
lines simply because they are made of oats,"
says Merrill Lynch food analyst William
Maguire.
63
Discussion exploitation
The oat-bran craze has cost the world's largest
cereal maker market share. The company's
president quit suddenly. And now Kellogg is
indefinitely suspending work on what was to
be a 1 billion cereal plant. The company said
it was delaying construction because of
current market conditions. But the Memphis,
Tenn., facility wasn't to begin turning out
product until 1993, so the decision may
reveal a more pessimistic long-term outlook as
well. Kellogg, which hasn't been as
successful in capitalizing on the public's
health-oriented desire for oat bran as rival
General Mills Inc., has been losing share in
the 6 billion ready-to-eat cereal market.
Kellogg's current share is believed to be
slightly under 40 while General Mills'
share is about 27. Led by its oat-based
Cheerios line, General Mills has gained an
estimated 2 share so far this year, mostly at
the expense of Kellogg. Each share point is
worth about 60 million in sales. Analysts say
much of Kellogg's erosion has been in such
core brands as Corn Flakes, Rice Krispies and
Frosted Flakes, which represent nearly
one-third of its sales volume. Kellogg is so
anxious to turn around Corn Flakes sales
that it soon will begin selling boxes for as
little as 99 cents, trade sources say.
"Cheerios and Honey Nut Cheerios have eaten away
sales normally going to Kellogg's corn-based
lines simply because they are made of oats,"
says Merrill Lynch food analyst William
Maguire.
64
Discussion exploitation
8gtgtgt The oat-bran crazee190 has coste189 the
world's largest cereal maker market share. 42gtgtgt
much of Kellogg's erosione204 has been in such
core brands as Corn Flakes, Rice Krispies and
Frosted Flakes, 49gtgtgt that it soont207 will
begine33 sellinge34 boxes for as little as 99
cents, 54gtgtgt "Cheerios and Honey Nut Cheerios
have eatene36 away sales 55gtgtgt normally
goinge38 to Kellogg's corn-based lines
8
42
49
54
55
65
Discussion exploitation
8gtgtgt The oat-bran crazee190 has coste189 the
world's largest cereal maker market share. 42gtgtgt
much of Kellogg's erosione204 has been in such
core brands as Corn Flakes, Rice Krispies and
Frosted Flakes, 49gtgtgt that it soont207 will
begine33 sellinge34 boxes for as little as 99
cents, 54gtgtgt "Cheerios and Honey Nut Cheerios
have eatene36 away sales 55gtgtgt normally
goinge38 to Kellogg's corn-based lines
8
42
49
54
55
66
Discussion exploitation
8gtgtgt The oat-bran crazee190 has coste189 the
world's largest cereal maker market share. 42gtgtgt
much of Kellogg's erosione204 has been in such
core brands as Corn Flakes, Rice Krispies and
Frosted Flakes, 49gtgtgt that it soont207 will
begine33 sellinge34 boxes for as little as 99
cents, 54gtgtgt "Cheerios and Honey Nut Cheerios
have eatene36 away sales 55gtgtgt normally
goinge38 to Kellogg's corn-based lines
8
42
49
54
55
67
Discussion exploitation
8
42
49
54
55
68
Conclusions

• Claims that a correlation exists between an
  RST/VT-type discourse structure and the temporal
  relations
• The claims have been supported by a corpus study
• Applicability
• manual development of gold corpora annotated
  for time
• discourse parsing, when temporal relations are
  known
• time tracking, when the discourse structure is
  known
• chronological focussed summarisation

69
Acknowledgements

• We are grateful to Daniel Marcu for making
  available the corpus of RST trees used in this
  study.
• This work has been partially supported by the
  CEEX research grant ROTEL-29 of AMCSIT (Romanian
  Ministry of Education and Research code
  PC-D03-PT00-205).

THANK YOU!